基于YOLOv5与背景优化的静态手势识别算法  

Static Hand Gesture Recognition Algorithm Based on Deep Learning and Background Optimization

作　　者：李书娴 王宇翔 赵雪峰 仲兆满 LI Shuxian;WANG Yuxiang;ZHAO Xuefeng;ZHONG Zhaoman(School of Information Engineering,Jiangsu Colledge of Finance&Accounting,Lianyungang 222061,China;College of Computer Engineering,Jiangsu Ocean University,Lianyungang 222005,China)

机构地区：[1]江苏财会职业学院信息工程学院,江苏连云港222061 [2]江苏海洋大学计算机工程学院,江苏连云港222005

出　　处：《山西大学学报(自然科学版)》2025年第1期180-191,共12页Journal of Shanxi University(Natural Science Edition)

基　　金：国家自然科学基金(72174079);江苏省“青蓝工程”优秀教学团队项目(2022-29);江苏海洋大学“研究生科研与实践创新计划项目”(KYCX2023-79)。

摘　　要：手势识别是人机交互中的关键技术。传统实时手势识别模型对光照变化、复杂背景等干扰因素适应性不强,所用分类数据集仅包含特定手势,在实际应用中泛化能力不足。针对以上问题,提出背景优化的二阶段静态手势识别算法。在检测阶段,采用YOLOv5s(You Only Look Once version 5 small)作为检测网络,利用其定位能力快速检测手部位置。在识别阶段,首先,利用背景与传感器热噪声对分类数据集进行增强,设计背景优化预处理算法,提升模型对复杂背景的适应性;然后,将VGG-16(Visual Geometry Group-16)作为识别网络的原型,增加归一化层并替换激活函数以加速收敛并防止过拟合。实验中,模型可以在多种干扰下提取图像特征,准确率达到97.9%,F1值达到92.3%。实验结果表明,模型对复杂场景的适应能力高于经原始分类数据集训练后的传统模型,具有更高的实际应用价值。Gesture recognition is a key technology in human-computer interaction.Traditional real-time gesture recognition models are less adaptive to disturbing factors such as illumination changes and complex backgrounds,and the classification dataset used only contains specific gestures,which is not enough for generalization in practical applications.To address these problems,a background-optimized two-stage static gesture recognition algorithm was proposed.In the detection phase,YOLOv5s(You Only Look Once version 5 small)was used as the detection network to quickly detect the hand position using its localization capability.In the recognition phase,firstly,the background and sensor thermal noise was used to enhance the classification dataset,and the background optimization preprocessing algorithm was designed to improve the model's adaptability to complex backgrounds;then,VGG-16(Visual Geometry Group-16)was used as the prototype of the recognition network,adding normalization layers and replacing the activation function to accelerate convergence and prevent overfitting.In the experiments,the model can extract image features under various disturbances with 97.9%accuracy and 92.3%F1 value.The experimental results show that the model is more adaptive to complex scenes than the traditional model trained by the original classification dataset,and has higher practical application value.

关 键 词：手势识别 YOLOv5 复杂背景 实时识别 

分 类 号：TP391.4[自动化与计算机技术—计算机应用技术]